CoT-style data for SFT and GRPO

[Shengqi77]

Hello~ I noticed you constructed 90K CoT-style data samples for both SFT and GRPO training. Could you clarify:

1. How many samples were used in each phase respectively?

2. Is the CoT annotation unnecessary during the GRPO stage?"

[Mountchicken]

HI @Shengqi77

1. For the SFT stage, we trained for two epochs using the entire 90K dataset. For the GRPO stage, we reused the same dataset for training, but we didn’t complete the full training, instead, we stopped once the reward had stabilized.

2. Yes, during the GRPO stage, the CoT content is not used for supervision.